Your alt text passes automated checks. That doesn't mean it's any good.

This post details the development of an alt text plugin for the GitHub Accessibility Scanner. It introduces five deterministic rules for identifying common alt text issues (e.g., filenames, placeholders, generic descriptions, repetition) based on string analysis alone. It also describes an opt-in, model-backed rule that uses page context (headings, captions, surrounding prose, link status) and image content to provide more nuanced feedback. Key technical details include using Playwright for DOM traversal, a layout-based check for repeated alt text using bounding box comparisons, and a decision procedure for the model-backed rule to ensure consistent verdicts. Privacy considerations like URL redaction and untrusted input handling are also discussed.